train: preflight -- download_checkpoint prefix/landing-dir fix (found by P1 v2)

train/preflight.py

@@ -206,6 +206,13 @@ ounce100m_credentials.install()
 from huggingface_hub import HfApi
 api = HfApi(); tok = os.environ["HF_TOKEN"]
 repo = "Cion-lab/ounce100m-ckptbench-DELETEME"
+# Start from an empty repo: LFS dedupes identical bytes, so pushing into a repo that already holds this
+# mock would time a no-op and report it as upload speed. Same lesson as the rehearsal's preclean.
+try:
+    api.delete_repo(repo_id=repo, repo_type="dataset", token=tok)
+    print("preclean: deleted stale ckptbench repo", flush=True)
+except Exception as e:
+    print("preclean:", type(e).__name__, str(e)[:100], flush=True)
 print("ensure_repo:", json.dumps(hubckpt.ensure_repo(repo, api, tok)), flush=True)
 d = "/kaggle/working/ckptbench/checkpoint-1"
 os.makedirs(d, exist_ok=True)
@@ -228,8 +235,9 @@ try:
     # second, independent proof: pull it back into a clean directory and compare hashes
     back = "/kaggle/working/ckptbench/restored"
     got = hubckpt.download_checkpoint(repo, "ckpt/checkpoint-1", back, api, token=tok)
-    same = json.load(open(os.path.join(back, "cursor.json"))) == {"step": 1,
-                                                                  "samples_consumed": 381500}
+    same = json.load(open(os.path.join(got["dir"], "cursor.json"))) == {
+        "step": 1, "samples_consumed": 381500}
+    print("download landed in", got["dir"], "files", got["files"], flush=True)
     # The repo is a timing rig, not an artifact: leaving a 1.7 GB public blob invites a future session
     # to mistake it for a checkpoint. Everything measurable is in the log by this point.
     api.delete_repo(repo_id=repo, repo_type="dataset", token=tok)